JPH0744188A - Speech recognition device - Google Patents

Abstract

PURPOSE:To detect a similar word with high precision by artificially generating an input pattern from a standard model, performing recognition by using the input pattern, and regarding found similarity as similarity between words. CONSTITUTION:In similar word detection mode, a detection/recognition changeover switch 19 is connected to a dummy input pattern generation part 13 and a detection/recognition changeover switch 20 is connected to a similar word determination part 15. The dummy input pattern generation part 13 reads a phoneme sequence corresponding to a word out of a word dictionary part 12 and reads the mean vector of standard models corresponding to respective phonemes out of a standard model storage part 11 to generate a dummy input pattern. The pattern is inputted to a recognition part 14 and the similarities to the words in the word dictionary part are found by using the standard models in the standard model storage part 11. A similar word determination part 15 outputs the word and similarity as information on a similar word and informs a user of it when the similarity is larger than a previously set threshold value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device suitable for fields requiring high recognition performance.

[0002]

2. Description of the Related Art Conventionally, in speech recognition, a standard model created from learning data uttered in advance is used to find the similarity between them and a speech pattern found from unknown speech, and a category giving the maximum similarity is recognized. The resulting method is commonly used. When the category is a word, if there are words with similar utterances among the words to be recognized, such as "Chiba" and "Shiga," when these sounds are input during recognition. However, there is a problem that misrecognition is likely to occur. This problem was especially noticeable in large vocabulary speech recognition, which targets many vocabulary words. In the large vocabulary speech recognition method, a unit smaller than a word such as a phoneme is generally used as a recognition unit. Below, "phoneme"
Means not only the minimum basic unit of speech in the phonological sense, but also a wider range of units of speech, including syllables and concatenation of multiple phonemes. As a method of using a phoneme as a recognition unit, for example, Watanabe, Yoshida, Koga et al., IEICE Transactions D-II.
Vol. J72-D-II No. 8 The method described in the paper “Large word speech recognition using HMM with semisyllabic units” (hereinafter referred to as Document 1) published on pages 1264-1269 of August 1989 can be mentioned. In this method, a standard model is created for each semi-syllable (hereinafter referred to as a phoneme), which is a kind of phoneme, using a plurality of learning data uttered in word units. At the time of recognition, a word-by-phoneme dictionary is used to combine standard models to create a word-by-word model, and this word model is used to recognize unknown word speech.

As a method for dealing with the problem of erroneous recognition due to similar words as described above, in addition to improving the accuracy of the standard model and the recognition method, similarity between words to be recognized in advance before recognition is performed. Find a combination of words with high similarity,
There is a method of notifying it to the user and excluding a part or all of the words in the combination from the recognition target or replacing the word with another word. An example of such a method of detecting similar words is described in Japanese Patent Publication No. 4-62595 (hereinafter referred to as Document 2). Reference 2
Then, prepare tables that define the distance between vowels and the distance between consonants, and make correspondence in syllable units for every two arbitrary words of the recognition target, and the distance between corresponding syllables. From the above two tables and using them, the similarity between words is checked.

[0004]

In the method of Document 2 described above, since the similarity between words is verified without depending on the recognition unit or the standard model, there is a tendency different from erroneous recognition that occurs during recognition. Is likely to detect similar words in
In addition, there is a problem that it is not possible to deal with cases where the vowels are similar to each other due to vowel deformation such as lengthening of vowels and vowel devoicing.

According to the present invention, in speech recognition using a phoneme as a recognition unit, a pseudo input pattern is created from a standard model, recognition is performed using this input pattern, and the obtained similarity is calculated as the similarity between words. The purpose of this is to detect similar words with high accuracy.

[0006]

A speech recognition apparatus of the present invention defines a feature analysis unit for analyzing a voice signal and outputting a feature vector time series, and a probability distribution in which the appearance probability for the feature vector is arbitrary. Standard model storage unit that creates and stores a standard model expressed as a network in a phoneme unit in advance, a word dictionary unit that stores phoneme information that constitutes a word that is the target of speech recognition, and a similar word A detection / recognition switch for switching between detection and recognition modes, a pseudo input pattern creation unit that creates a pseudo input pattern from phoneme information for any word in the word dictionary unit and the standard model in the similar word detection mode, In the similar word detection mode using the standard model and the phoneme information stored in the word dictionary, the pseudo input pattern and the recognition mode are detected. The recognition unit recognizes the feature vector time series output from the feature analysis unit and obtains the similarity with the word that is the target of the voice recognition, and the similarity output from the recognition unit. From the similar word determination unit that outputs a word having a high similarity from the words that are the target of the voice recognition as a similar word, the similarity between the words output from the recognition unit, the voice signal and the target of the voice recognition. A recognition result determination unit that outputs a recognition result by checking the similarity with a word, and a result display unit that displays the recognition result in the similar word detection mode and the recognition result in the recognition mode. Alternatively, transition probabilities are defined for the transitions between states in the standard model stored in the standard model storage unit, and the pseudo input pattern creation unit considers the transition probabilities. Characterized by creating a pattern.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing a first embodiment of the present invention. The standard model storage unit 11 stores in advance a standard model P _k (k = 1, 2, ... K, K is the number of types of phonemes) in units of phonemes. As the standard model P _k , the HMM described in Document 1 can be used and can be created by the learning method described in Document 1. The HMM is a kind of state transition network, and each state i (i = 1, 2, ... I ^k , I ^k is the number of states of the standard model P _k ) has a state transition probability.

[0009]

[Equation 1]

And the feature vector appearance probability are defined. The Gaussian distribution is used as the feature vector appearance probability.

[0011]

[Equation 2]

Can be used. Word dictionary section 12
Is a word W _m to be recognized (m = 1, 2, ...,
Phoneme information forming M and M is the vocabulary number) is stored in advance. As the phoneme information, a phoneme string of the word W _m

[0013]

[Equation 3]

Can be used.

Next, the operation of this embodiment will be described with reference to FIGS. 1 and 2.

In the similar word detection mode, the detection / recognition changeover switch 19 is connected to the pseudo input pattern creating section 13, and the detection / recognition changeover switch 20 is connected to the similar word determining section 15.

The pseudo input pattern creating unit 13 uses the phoneme string for the word W _m from the word dictionary unit 12.

[0018]

[Equation 4]

Is read, the average vector of the standard model corresponding to each phoneme is read from the standard model storage unit 11,
A pseudo input pattern is created (step A ₁ ). phoneme

[0020]

[Equation 5]

The standard model corresponding to

[0022]

[Equation 6]

And the mean vector contained therein is

[0024]

[Equation 7]

Then, as the pseudo input pattern V _m for the word W _m ,

[0026]

[Equation 8]

Then, a pattern in which average vectors are arranged in a line can be used. When arranging the average vectors, 1
It is possible to arrange multiple pieces instead of individual pieces.

The generated pseudo input pattern V _m is input to the recognition unit 14 and, using the standard model of the standard model storage unit 11, the word W _n (n = 1, 2, ...
, M) and the degree of similarity R _mn is obtained (step A ₂ ). As a method of calculating the pseudo degree, the recognition method described in Document 1 can be used. If the calculated similarity R _mn is larger than a preset threshold value, the pseudo word determination unit 15 outputs the words W _m , W _n and the similarity R _mn as information about similar words, and the result display unit 16 displays those information and informs the user (step A ₃ and step A ₄ ). Based on the information, the user can improve the recognition performance by excluding one or both of the similar words from the recognition target or replacing them with another word.

If the word W _n is the final word W _{M in} the word dictionary (step A ₅ ), that is, if the similarity verification with respect to the word W _m is executed for all the words in the word dictionary, the word W _n Similarly for _{m + 1} , steps A _{1 to} A ₅
To perform similarity verification. When the word W _m is the final word W _{M in} the word dictionary section (step A ₆ ), that is, when the verification of the similarity to all the words in the word dictionary section is completed, the processing is ended.

In the recognition mode, the detection / recognition changeover switch 19 is connected to the feature analysis section 17, and the detection / recognition changeover switch 20 is connected to the recognition result determination section 18.

In the characteristic analysis unit 17, Furui, 1985.
The unknown speech signal was converted into a feature vector time series using the method by the mel cepstrum as described in "Digital Speech Processing" published by Tokai University Press
In the recognition unit 14, the feature vector time series and the word W _n (n = 1, 2, ...
.., M) and the degree of similarity R _mn is obtained. The recognition result determination unit 18 selects an arbitrary number from the descending order of the _calculated similarity R _mn , outputs the value and the corresponding word as a recognition result, and the result display unit 16 displays the information.

FIG. 3 is a block diagram showing a second embodiment of the present invention.

Referring to FIG. 3, in the second embodiment of the present invention, the number of vectors between the standard model storage unit 11 and the pseudo input pattern creating unit 13 in the first embodiment of the present invention shown in FIG. The difference is that the decision unit 21 is added.

The operation of this embodiment is different from the operation of the first embodiment shown in FIG. 2 in the method of creating the pseudo input pattern in step A ₁ , and the other operations are the same. In the first embodiment, a pseudo input pattern is created by arranging a fixed number of average vectors, and the state transition probability in the standard model is

[0035]

[Equation 9]

Was not used. In the present embodiment, the vector number determination unit 21 uses the standard model storage unit 11 to calculate the state transition probability of the standard model.

[0037]

[Equation 10]

The number of average vectors read by and arranged using the values

[0039]

[Equation 11]

Determine. Number

[0041]

[Equation 12]

The following equation can be used to obtain

[0043]

[Equation 13]

The pseudo input pattern creating unit 13 creates a pseudo input pattern from the obtained number, the phoneme sequence in the word dictionary unit 12, and the average vector of the standard model in the standard model storage unit 11.

In addition to the second embodiment, as a method of using the state transition probability at the time of creating the pseudo input pattern, the phoneme information of the word W _m in the word dictionary unit 12 has a branch as shown in FIG. In the case of a string, for example, when expressing a word in which a vowel at the end may be unvoiced, a phoneme of a branch having a high state transition probability is used for creating a pseudo input pattern.

[0046]

As described above, in the similar word detection method according to the present invention, recognition is performed using an input pattern pseudo-created from a standard model, and the obtained similarity is regarded as the similarity between words. There is an effect that a similar word can be detected with high accuracy.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a flowchart in a similar word detection mode according to the embodiment of this invention.

FIG. 3 is a block diagram showing a second embodiment of the present invention.

FIG. 4 is a diagram showing an example of phoneme information of words in a word dictionary unit of the present invention.

[Explanation of symbols]

 11 standard model storage unit 12 word dictionary unit 13 pseudo input pattern creation unit 14 recognition unit 15 similar word determination unit 16 result display unit 17 feature analysis unit 18 recognition result determination unit 19, 20 detection / recognition switch 21 vector number determination unit

Claims (2)

[Claims] 1. A standard analysis model for expressing a feature vector time series by analyzing a voice signal and a network of states in which appearance probabilities for feature vectors are defined in the form of an arbitrary number of probability distributions. A standard model storage unit that creates and stores each phoneme in advance, a word dictionary unit that stores phoneme information that constitutes a word that is a target of speech recognition, and detection / recognition for switching between similar word detection and recognition modes. A changeover switch, a pseudo input pattern creation unit that creates a pseudo input pattern from the phoneme information for any word of the word dictionary unit and the standard model in the similar word detection mode, the standard model and the word dictionary unit. The phoneme information is used to detect the pseudo input pattern in the similar word detection mode, and the feature output from the feature analysis unit in the recognition mode. A recognition vector time series, and a recognition unit that obtains the degree of similarity with the word that is the target of the voice recognition; and the similarity that is output from the recognition unit, from the word that is the target of the voice recognition. A similar word determination unit that outputs a word with high similarity as a similar word, and a similarity determination between the voice signal and the word that is the target of the voice recognition is performed based on the similarity output from the recognition unit, and a recognition determination is performed. A speech recognition apparatus, comprising: a recognition result determination unit that outputs a recognition result; and a result display unit that displays the similar word in the similar word detection mode and the recognition result in the recognition mode. 2. A transition probability is defined for a transition between states in the standard model stored in the standard model storage unit, and the pseudo input pattern creation unit considers the transition probability to generate a pseudo input pattern. The voice recognition device according to claim 1, wherein the voice recognition device is created.